Compute now has a reference price, and it is live

When I joined LO:TECH back in June, one of the first things I took on was building a proper reference price for GPU compute. As of this week, our compute pricing feed is serving as the benchmark for the first ever multi-GPU index perpetual, the Super Intelligence Compute Index (SICI), which launched on Injective mainnet on October 26th.

The index perpetual on Injective, which trades with up to 10x leverage, is just the first product powered by our latest feed. The underlying per-chip layer gives us huge flexibility and scope to grow, whether someone needs a single-name H100 or B200 perpetual reference or a settlement source for a prediction market on rental rates, the data is ready to go.

There is still more to share soon, including a deeper look into our spot-to-on-demand ratio as a scarcity signal and our residual-based method for rolling index generations without relying on committee decisions. First I want to share the story of how this family of products came together, how the feed works, and where we are taking it next.

Why compute needed a reference price

The core premise was straightforward, pricing GPU compute has been a trend over the past year and several players have rushed to offer this service. However, executing this end-to-end was a real journey with plenty of highs and lows. The market is nascent and fragmented with lots of idiosyncrasies you need to be aware of when pricing.

An hour of Nvidia’s flagship H100 GPU gets bought and sold constantly across a fragmented landscape of marketplaces, each using different product names, regions, and billing conventions. Until recently, there was no single defensible reference price to tie it all together. Coming into LO:TECH, this felt like a natural extension of what the firm has always done for crypto assets through its trading and market making business. The challenge of collecting raw quotes, mapping them and defending the aggregation is fundamentally the same whether you are dealing with an exchange or a neocloud.

How the feed works

Building and owning this project meant tackling the technical pipeline head-on, and required plenty of help from our data engineering team. The feed tracks marketplace listings across our covered providers to generate an hourly reference price for each chip in dollars per GPU-hour. Every observation passes through rigorous validation and outlier controls before reaching the final price. Crucially, every record is stored with its raw payload alongside the normalized value, with zero aggregation at collection time. That determinism is crucial, it means six months down the line, any print can be regenerated from the archive with an identical result when someone trades against it.

The challenge of collecting raw quotes, mapping them and defending the aggregation is fundamentally the same whether you are dealing with an exchange or a neocloud.

Five chips, fixed weights

On top of these individual chip prices sits the index itself, covering five NVIDIA chips at fixed weights:

• H100 SXM (40%)

• B200 (25%)

• H200 (15%)

• A100 SXM (10%)

• RTX 4090 (10%)

The benchmark price is calculated on the hour every hour, with the 19:00 UTC print acting as the daily fixing, and the published level reflects a trailing seven-day average of the hourly basket. We deliberately chose fixed weights with our partner because an index that rebalances on its own schedule introduces artificial noise unrelated to compute prices, and a settlement reference should only move when the underlying market moves.

Getting here

This week’s success wasn’t linear though, and the broader AI landscape played a part in the ups and downs along the way. We’ve navigated potential sales leads and integration deals that ultimately fell through, but keeping full ownership of the product allowed us to stay focused on building something robust.

If you are exploring compute products, whether index perps, single-name contracts, or prediction markets, reach out to me directly at data@lo.tech and we will be happy to walk you through the feed and data architecture.